STOCK BROAD-INDEX TREND PATTERNS LEARNING VIA DOMAIN KNOWLEDGE INFORMED GENERATIVE NETWORK

International Journal of Artificial Intelligence and Applications (IJAIA), Vol.14, No.2, March 2023
DOI: 10.5121/ijaia.2023.14202 11
STOCK BROAD-INDEX TREND PATTERNS
LEARNING VIA DOMAIN KNOWLEDGE INFORMED
GENERATIVE NETWORK
Jingyi Gu, Fadi P. Deek and Guiling Wang
New Jersey Institute of Technology, Newark, NJ, USA
ABSTRACT
Predicting the Stock movement attracts much attention from both industry and academia. Despite such
significant efforts, the results remain unsatisfactory due to the inherently complicated nature of the stock
market driven by factors including supply and demand, the state of the economy, the political climate, and
even irrational human behavior. Recently, Generative Adversarial Networks (GAN) have been extended for
time series data; however, robust methods are primarily for synthetic series generation, which fall short
for appropriate stock prediction. This is because existing GANs for stock applications suffer from mode
collapse and only consider one-step prediction, thus underutilizing the potential of GAN. Furthermore,
merging news and market volatility are neglected in current GANs. To address these issues, we exploit
expert domain knowledge in finance and, for the first time, attempt to formulate stock movement prediction
into a Wasserstein GAN framework for multi-step prediction. We propose Index GAN, which includes
deliberate designs for the inherent characteristics of the stock market, leverages news context learning to
thoroughly investigate textual information and develop an attentive seq2seq learning network that captures
the temporal dependency among stock prices, news, and market sentiment. We also utilize the critic to
approximate the Wasserstein distance between actual and predicted sequences and develop a rolling
strategy for deployment that mitigates noise from the financial market. Extensive experiments are
conducted on real-world broad-based indices, demonstrating the superior performance of our architecture
over other state-of-the-art baselines, also validating all its contributing components.
KEYWORDS
Stock Market Prediction, Generative Adversarial Networks, Deep Learning, Finance
1. INTRODUCTION
The stock market is an essential component of a broad and intricate financial system, and stock
prices reflect the dynamics of economic and financial activities. Predicting future movements of
either individual stocks or overall market indices is important to investors and other market
players [1], which requires significant efforts but lacks satisfactory results. Conventional
approaches vary from fundamental and technical analysis to linear statistical models, such as
Momentum Strategies and Autoregressive Integrated Moving Average (ARIMA), which capture
simple short-term patterns from historical prices. With the tremendous power and success in
exploring the nonlinear relationship and dealing with big data, machine learning and neural
networks are increasingly utilized in stock movement prediction and have shown better results in
prediction accuracy over traditional methods [2].
However, stock prices, especially broad-based indices (such as Dow30 and S&P 500), carry
inherent stochasticity caused by crowd emotional factors (such as fear and greed), economic
factors (such as fiscal and financial policy and economic health), and many other unknown

12
factors [3]. Thus, existing models suffer from a lack of robustness, and the prediction accuracy is
still far from being satisfactory. Additionally, unlike many other tasks, the prediction accuracy of
broad-index return cannot keep increasing, in theory, because market arbitrage opportunities will
degrade accuracy over time. Naturally, an effective trading strategy that is likely to generate
profit will be adopted by an increasing number of investors; eventually, such a strategy may no
longer perform as well considering that the market is composed of its collective participants.
Thus, in practice, if a certain method is capable of achieving over 60% accuracy using a dataset
of a lengthy period, encompassing a financial crisis, it is then considered powerful enough.
Therefore, we aim to provide a framework that performs with accuracy, over time, and in
different market conditions.
Recently, a powerful machine learning framework named Generative Adversarial Networks
(GAN) [4] achieved noteworthy success in various domains. Distinct from single neural networks,
GAN trains two models contesting with each other simultaneously, a generator capturing data
distribution and a discriminator estimating the probability of a sample from training data instead
of a generator. Having achieved promising performance in image generation, researchers have
extended GANs to the field of time series for applications in synthetic time series data generation.
While aiming to simulate characteristics in sequences, these efforts have rarely been practical
enough to predict future patterns. Although some advanced methods have been proposed to
forecast stock pricing by building an architecture of GAN conditioned on historical prices [5, 6],
they have two main deficiencies: 1) Their implemented vanilla GAN suffers from mode collapse
problem [7] and creates samples with low diversity. Consequently, most predictions tend to be
biased towards an up-market. 2) Existing models only predict one future step, which is then
concatenated with historical prices to generate fake sequences. As a result, the only difference
between the actual and generated sequence is the last element, and thus the discriminator can
easily ignore it. Such methods largely underutilize GAN’s potential, which can create sequences
with periodical patterns of the stock market. We aim to address these issues by fully realizing the
potential of GAN to solve the stock index prediction problem.
In addition to historical prices, other driving factors such as emerging news and market volatility
should also be exploited in market prediction. An increasing amount of research results has
shown that the stock market is highly affected by significant events [8]. With the popularity of
social media, frequent financial reporting, targeted press releases, and news articles significantly
influence investors’ decisions and drive stock price movement. Natural Language Processing
(NLP) techniques have been utilized to analyze textual information to forecast stock pricing. In
addition to news, the volatility index, derived from options prices, contains important indicators
on market sentiment [9, 10, 11]. There are some common similarity between options and
insurance. For example, put options can be purchased as insurance, or a hedge, against a sudden
stock price drop. A price increase in put options with the same underlying stock price indicates
that the market expects a high probability of price drop, just as an increased flood insurance price
imply a higher probability of flood though people may still choose to stay in their homes (i.e., to
keep the stock position unsold). The volatility index, which is derived from the prices of index
options with near-term expiration dates, is an important leading indicator of the stock market but
is often neglected in the GAN models for predicting the stock market. We incorporate both news
and volatility index in our proposed prediction model.
Because existing efforts predominantly rely on historical prices and neglect the inherent
characteristics of the stock market, there is a need to better address the issues and yield higher
prediction accuracy. Thus, we propose the following strategies which employ domain knowledge
in financial systems: 1) formulating stock movement to a multiple-step prediction problem; 2)
exploiting Wasserstein GAN to avoid mode collapse that tends to predict an up-market, and
improve stability by prior noise; 3) leveraging volatility index to capture the market sentiment of

13
seasoned investors; 4) designing news context learning to fully explore textual information from
public news and events; and 5) adopting a rolling strategy for deployment to mitigate noise from
uncertainty in the financial market, and generate the prediction with horizon-wise information.
Specifically, our proposed IndexGAN contains both a generator and a critic to predict the stock
movement. Inside the generator, a news context learning module extracts latent word
representations from GloVe embedding matrix of news for two purposes. Firstly, it fully explores
context information from the news. Secondly, it strategically shrinks the size of the embedding
matrix to a comparable extent with price features, considering the size of GloVe embedding
matrix is several orders of magnitude larger than market-based features and prior noise, and
simply concatenating them together will let the following seq2seq learning entirely concentrate
on news and mask the impact of other features. An attentive seq2seq learning module captures
temporal dependence from historical prices, volatility index, various technical indicators, prior
noise, and latent word representations. The generator with both modules predicts the future
sequence. The critic (discriminator in GAN) is designed to approximate the Wasserstein distance
between real and generated sequences, which is incorporated into an adversarial loss for
optimization. During the training process, besides adversarial loss, two additional losses are
added to improve the similarity between real and fake distribution efficiently and avoid biased
movement.
Accurate predictions of an index return enable hedging against black swan events and thus can
significantly reduce risk and help maintain market liquidity when facing extreme circumstances,
thereby contributing to financial market stabilization. IndexGAN is implemented on broad indices.
To understand the complexity and dynamics of the real-world market index, we select a period of
high volatility that covers the 2009 financial crisis and the subsequent recovery for training and
testing of IndexGAN. Experiment results show that the model boosts the accuracy from around 58%
to over 60% for both indices (Dow Jones Industrial Average (DJIA) and S&P 500) compared to
the previous state-of-the-art model. This improvement demonstrates the effectiveness of the
proposed method. The major contributions of this work are summarized as follows:
1. To the best of our knowledge, ours is a first attempt at formulating stock movement
predictions into a Wasserstein generative adversarial networks framework for multi-step
prediction. The new formulation enhances the robustness by prior noise and provides
reasonable guidance on market timing to investors through the use of multi-step
prediction.
2. IndexGAN consists of particularly crafted designs that consider the inherent
characteristics of the stock market. It includes a news context learning module, features
of volatility index and multiple empirically valid technical indicators, and a rolling
deployment strategy to mitigate uncertainty and generate prediction with horizon-wise
information.
3. We conduct extensive experiments on two stock indices that attract the most global
attention. IndexGAN achieves consistent and significant improvements over multiple
popular benchmarks.
2. RELATED WORKS
2.1. Stock Market Prediction
Stock movement prediction has drawn much industry attention for its potential to maximize profit,
as has the intricate nature of this problem to academia. Traditional methods fall into two
categories: fundamental analysis and technical analysis [12]. Fundamental analysis attempts to
evaluate the fair and long-term stock price of a firm by examining its fundamental status and
comparative strength within its industry as well as the overall economy such as revenues,

14
earnings, and account payables [13, 14]. Technical analysis, also known as chart analysis,
examines price trends and identifies price patterns [15]. There are numerous technical indicators,
such as moving average and Bollinger Bands, that guide traders on buying and selling trends.
Linear models, such as ARIMA [16] and Generalized Autoregressive Conditional
Heteroskedasticity (GARCH) [17], have also been utilized in stock prediction. With the
development of deep learning, many deep neural networks have been employed in the finance
area, especially for stock prediction [18]. Qin et al. [19] propose a dual-stage attention-based
Recurrent Neural Network (RNN) model to predict stock trends. Feng et al. [20] further consider
employing the concept of adversarial training on Long Short-Term Memory (LSTM) [21] to
predict stock market pricing. Ding et al. [22] enhance the transformer with trading gap splitter
and gaussian prior for stock price prediction.
As the outreach of digital and social media grow enabling real-time streaming of financial news,
their impact on the stock market is explored. Existing work for textual stock analysis can be
classified into sentiment analysis and context-aware approaches. Sentiment analysis is leveraged
to extract emotions from text and generate sentiment scores as features [23], although this often
ignores useful textual information. Zhang, Fuehres, and Gloor [24] use mood words as emotional
tags to classify the sentiment of tweets. De Fortuny et al.[25] select sentiment words from the
news by TF-IDF and calculate sentiment polarity and subjectivity as sentiment features for
Support Vector Machine (SVM) in making predictions. Zhao et al. [26] employ LDA to filter
microblogs and create a financial sentiment lexicon to extract public emotions. On the other hand,
context-aware approaches create embeddings from news and feed them into complicated
structures to explore context as much as possible [27, 28, 29], which often suffer from the
overfitting problem. Lee and Soo [30] process financial news with word2vec to create input for
recurrent Convolutional Neural Network (CNN). Minh et al. [31] implement Stock2Vec and two-
stream Gated Recurrent Unit (GRU) models to create embeddings from financial news for stock
price classification. Li et al. [32] propose using Graph Convolutional Network to represent
connection given news embeddings and predict overnight movement. Both methods are widely
used for leveraging news in stock market prediction.
Although existing efforts have been successful in achieving their stated goals, market risk is still
largely not adequately addressed. Hence, we aim to incorporate market volatility into stock
prediction.
2.2. Generative Adversarial Nets on Time Series
More recently, there are emerging efforts in using GAN on time series data in the domain of
video generation [33] in medicine [34], text generation [35], biology [36], finance [37], and
music [38]. These GANs aim to generate a synthetic sequence that retains the characteristics of
the original sequence. For instance, Yu et al. [39] introduce SeqGAN, which extends GANs with
reinforcement learning to generate sequences of discrete tokens. Yoon et al. [40] propose
TimeGAN for realistic time-series generation, combining an unsupervised paradigm with
supervised autoregressive models. These methods are used for tasks of enriching limited real-
world datasets [41], data imputation [42], and machine translation [43]. However, stock market
predictions are attempts to discover future patterns instead of simulating original time series and
thus cannot employ most existing efforts on time series GANs.
Some recent approaches implement GANs to predict stock trends. Zhang et al. [44] employ
LSTM as a generator and MLP as a discriminator to forecast stock pricing. Faraz and
Khaloozadeh [6] enhance GAN with wavelet transformation and z-score method in the data
preprocessing phase. Sonkiya et al. [45] generate binary sentiment score by FinBERT and
employ GAN for stock price predictions. Such methods adapt Vanilla GAN which often suffers

15
from mode collapse, and neglect possible driving factors such as complex public emotions,
financial policies, and events, which limits the ability to achieve superior performance. In this
work, we use GAN with Wasserstein distance to design a robust architecture enabling us to
capture the characteristics of the stock market.
3. PROBLEM FORMULATION
We aim to develop a model that accurately predicts whether a stock index is up or down the next
day. If the close price of the next day is higher than that of the current day, we have an up day
and vice versa. We select two different types of features: market-based features read derived from
the stock market data and news-based features derived from raw news headlines. Market-based
features are listed in Table 1. Specifically, we retrieve the open, high, low, and close prices of the
stock index on trading days. We also deliberately select nine empirically useful technical
indicators, calculate their values based on the price data, and employ them as features. In addition,
we employ the market volatility data, i.e., VIX, if we predict the S&P 500 index and VXD if we
predict the Dow Jones Index.
Let 𝒮 = {𝑠𝑡}𝑡=1
𝑛
∈ ℝ𝑛×𝑝
be a matrix which represents the stock dataset with 𝑛 time points, where
𝑠𝑡 is a vector with 𝑝 market-based features. Letℋ = {ℎ𝑡}𝑡=1
𝑛
∈ ℝ𝑛×𝑘
be a matrix which denotes
textual dataset, we assume that each vector ℎ𝑡contains a number of 𝑘 raw news headlines at the
time step 𝑡. Let 𝑧 denote the prior noise which is assumed to be drawn from Gaussian distribution.
We concatenate stock features and headlines together. Then we denote our training data as 𝒟 =
{(𝑠𝑡,ℎ𝑡,𝑧)}𝑡=1
𝑛
. To investigate the temporal dependency of sequential data, we set the length of
historical sequence as 𝑤 and the length of future sequence as 𝑞. This means that we use the past
𝑤days’ data to predict the next 𝑞 days sequence in the inference phase. Suppose the current time
is day 𝑇 , given the trained model, we use 𝑋 = {(𝑠𝑡,ℎ𝑡,𝑧)}𝑡=𝑇−𝑤+1
𝑇
as input and generate
𝑐̂ = {𝑐̂𝑡}𝑡=𝑇+1
𝑇+𝑞
, where 𝑐̂𝑡 is the predicted return based on close price. From the generated
sequence, we can determine whether day 𝑇 + 1 is up or down.
Table 1. Market-based features. 𝑠𝑡 denotes corresponding features before processing. BOLU and BOLD
are upper and lower Bollinger Bands.
Type Features Calculation
Price Open, Close, High, Low 𝑠𝑡
𝑠𝑡−1
− 1
Volatility VIX, VXD
Technical
Indicator
RSI -1 if 𝑠𝑡 ≤ 20, 1 if 𝑠𝑡 ≥ 80, 0 otherwise
MACD DIFF min-max scaling
Upper Bollinger band indicator if 𝑐𝑙𝑜𝑠𝑒𝑡 > 𝐵𝑂𝐿𝑈, 0 otherwise
Lower Bollinger band indicator if 𝑐𝑙𝑜𝑠𝑒𝑡 < 𝐵𝑂𝐿𝐷, 0 otherwise
EMA: 5-day 𝑠𝑡
𝑐𝑙𝑜𝑠𝑒𝑡
− 1
SMA: 13-day, 21-day, 50-day, 200-day
4. METHODOLOGIES
We elect to use GAN, specifically IndexGAN, for stock market prediction. The architecture of the
model is shown in Figure 1. In general, GAN consists of two models: a generator that learns
distribution over real data, and a discriminator that distinguishes between real data and fake
output from the generator. For the purpose of controlling on modes of data being generated,
conditional GAN [46] was proposed by conditioning the model to additional information. In our
model, stock and news features are used as additional information. The goal is to use training
data 𝒟 to learn a feature map (generator) 𝒢 which best generates sequences of future stock prices

16
in the GAN framework. WGAN [47], which uses Wasserstein distance to measure the difference
between the real data distribution and generated data distribution, is chosen for its stability.
4.1. Generator
Generator 𝒢 with parameter set Θ𝑔 takes input 𝑋 and aims to build a fake sequence 𝑐̂ whose
distribution is as close as possible to the real sequence 𝑐 = {𝑐𝑡}𝑡=𝑇+1
𝑇+𝑞
. Our generator has two
parts: news context learning and attentive seq2seq learning.
4.1.1. News Context Learning
To convert the text from news into word vectors, word2vec is applied to encode the textual input
as a continuous distributed representation. GloVe, one of the most popular word representation
techniques in word2vec, is chosen to be employed in our model. Compared with
generalword2vec models, GloVe not only considers words and related local context information
but also incorporates aggregated global word co-occurrence statistics from the corpus to create an
embedding matrix. Given a paragraph ℎ𝑡 at time step 𝑡 consisting of 𝑘 news headlines, each word
in the headlines is mapped to a vector in 𝑚 dimensions. The process is formulated as:
ℎ′
𝑡 = 𝑓𝐺𝑙𝑜𝑉𝑒(ℎ𝑡) (1)
where 𝑓𝐺𝑙𝑜𝑉𝑒 is the mapping function from GloVe to word vectors, ℎ′
𝑡 ∈ ℝ𝑘×𝑙×𝑚
represents the
word embedding matrix, and 𝑙 is the maximum length of headlines in the textual dataset ℋ. Since
the length of words in different headlines may be different, we add zero paddings to the end of
word vectors from GloVe if the length is shorter than 𝑙.
Then a block of nonlinear layers for the embedding matrix is designed. The first component in a
block is a fully connected layer. Next, batch normalization is applied to accelerate the training
process by normalizing the features in the embedding matrix. In addition, since batch
normalization is shown effective in preventing the mode collapse problem, we apply it to all
blocks except the last one. The third layer is LeakyReLU activation which avoids saturation in
the learning process and dying neurons, except that the last block uses tanh activation function
for the output layer. We feed the embedding matrix ℎ′
𝑡into a number of 𝑚 successive blocks
formulated as follows:
ℎ𝑡
1
= 𝜙(𝑓𝐵𝑁(𝑊𝑏1ℎ𝑡
′
+ 𝑏𝑏1)) (2)
ℎ𝑡
2
= 𝜙(𝑓𝐵𝑁(𝑊𝑏2ℎ𝑡
1
+ 𝑏𝑏2)) (3)
…
ℎ𝑡
̃ = 𝑡𝑎𝑛ℎ(𝑊𝑏𝑚ℎ𝑡
𝑚−1
+ 𝑏𝑏𝑚) (4)
where ℎ𝑡
𝑖
,𝑖 = 1, … , 𝑚 is the output from 𝑖𝑡ℎ
block and input for 𝑖 + 1𝑡ℎ
block, ℎ𝑡
̃ ∈ ℝ𝑔
is the
latent word representation with 𝑔 dimensions, 𝜙 is the LeakyReLU activation function, 𝑊𝑏𝑖 and
𝑏𝑏𝑖 are parameters in a fully connected layer

17
Figure 1 The Proposed IndexGAN Structure
4.1.2. Attentive Seq2seq Learning
Stock data shows seasonality. For example, a high open price on Monday morning frequently
ends up with an up day on Friday. To adaptively capture the temporal relationship and generate
sequences, we employ an encoder-decoder with the attentive mechanism as the sequence-to-
sequence learning. To improve the robustness of prediction, we add prior noise as an input
feature to simulate the stochasticity nature of stock prices. In addition, among market-based
features, Exponential Moving Average (EMA) and Simple Moving Average (SMA) over
different periods of time (5, 13, 21, 50, and 200 days) are used to capture temporal information
across multi-scale time spans.
The encoder extracts the attentive variable 𝑎. Its input is the sequence of a concatenation of
features 𝑥𝑡containing prior noise 𝑧, market-based feature 𝑠𝑡, and latent word representation ℎ𝑡
̃ :
𝑥𝑡 = [𝑧, 𝑠𝑡,ℎ𝑡
̃ ] (5)
We adopt the simple yet effective GRU to obtain dynamic changes from historical sequence
{𝑥𝑡}𝑡=𝑇−𝑤+1
𝑇
:
ℎ𝑡
𝑒
= 𝑓
𝑒(𝑥𝑡,ℎ𝑡−1
𝑒 ) (6)
where 𝑓
𝑒 represents a GRU layer, and ℎ𝑡−1
𝑒
denotes the hidden state at each time step.
To adaptively learn the different contributions of input features in the different historical time
steps, we implement an attention layer to aggregate information of hidden state from GRU:
𝑎𝑡 = 𝑣⊤
𝑡𝑎𝑛ℎ(𝑊
𝑎ℎ𝑡
𝑒
+ 𝑏𝑎),𝑎𝑡
̃ =
𝑒𝑥𝑝(𝑎𝑡)
∑ 𝑒𝑥𝑝(𝑎𝑡)
𝑇
𝑡=𝑇−𝑤+1
(7)
𝑎 = ∑ 𝑎𝑡
̃ ℎ𝑡
𝑒
𝑇
𝑡=𝑇−𝑤+1
(8)

18
where 𝑎𝑡
̃ denotes contribution score for hidden state ℎ𝑡
𝑒
at time𝑡, 𝑣, 𝑊
𝑎 and 𝑏𝑎are parameters, and
𝑎 is the attentive representation which encloses periodical information of the entire sequence.
The decoder aims to extract patterns from the attentive representation 𝑎 and generate the fake
sequence 𝑐̂ ∈ ℝ𝑞
in the future 𝑞 time steps. We also adopt GRU as the decoder. The hidden state
in the decoder ℎ𝑇
𝑑
is transformed from the last hidden state in encoder ℎ𝑇
𝑒
by a fully connected
layer. Then the attentive representation 𝑎 is presented as input of the GRU. The process is
computed as follows:
ℎ𝑇
𝑑
= 𝑡𝑎𝑛ℎ(𝑊𝑑ℎ𝑇
𝑒
+ 𝑏𝑑) (9)
𝑐̂ = 𝑓
𝑓𝑐 (𝑓𝑑(𝑎, ℎ𝑇
𝑑
)) (10)
where 𝑓𝑑 denotes the GRU layer in decoder, 𝑊𝑑 and 𝑏𝑑 are weight and bias parameters, and 𝑓
𝑓𝑐
represents the prediction layer to generate the fake sequence 𝑐̂.
For training purposes, the next step is to feed both fake and real sequences into critic for further
optimization.
4.2. Critic
The critic is designed to approximate Wasserstein distance 𝑊(𝑃𝑟, 𝑃
𝑔) between real data
distribution 𝑃𝑟 and generated data distribution 𝑃
𝑔:
𝑊(𝑃𝑟, 𝑃
𝑔) = inf
γ∈Π(𝑃𝑟,𝑃𝑔)
𝐸(𝑥,𝑦)∼γ [||𝑥 − 𝑦||] (11)
where Π(𝑃𝑟, 𝑃
𝑔) denotes the set of all joint distribution 𝛾(x, y) whose marginals are 𝑃𝑟 and 𝑃
𝑔
respectively. The joint distribution 𝛾(𝑥, 𝑦) represents the mass caused by transporting 𝑥 to 𝑦when
transforming 𝑃𝑟 to 𝑃
𝑔 . Therefore, Wasserstein distance indicates the cost of such optimal
transportation problem in intuition [47,48].
Let 𝑐̃ denote the input sequence for either real or fake data. We use GRU to temporally capture
the dependency and a fully connected layer to pass the last hidden state, which is presented as:
𝑣 = 𝑓
𝑐(𝑐̃; 𝛩𝑐) = 𝑓
𝑣(𝑓𝐺𝑅𝑈(𝑐̃)) (12)
where 𝑓
𝑐 denotes the critic consisting of the GRU layer 𝑓𝐺𝑅𝑈 and the fully connected layer 𝑓
𝑐, 𝛩𝑐
represents the set of parameters in 𝑓
𝑐, and the output 𝑣 is the approximated distance value.
4.3. Optimization
As the infimum in 𝑊(𝑃𝑟,𝑃
𝑔) is highly intractable, Kantorovich-Rubinstein duality is used to
reconstruct the WGAN value function in critic and solve the optimization problem in practice,as
follows:
𝑚𝑖𝑛
𝒢
𝑚𝑎𝑥
𝑓𝑐∈ℱ
𝔼𝑐∼𝑃𝑟
[𝑓
𝑐(𝑐)] − 𝔼𝑐̂∼𝑃𝑔
[𝑓
𝑐(𝑐̂)] (13)
where 𝑃
𝑔 is implicitly defined by 𝑐̂ , ℱ is the set of 1-Lipschitz functions. The generator is
responsible for minimizing the value function in critic which is equivalent to minimize 𝑊(𝑃𝑟, 𝑃
𝑔),

19
so that the generated sequence is similar to the real sequence. The critic aims to find the optimal
value function to better approximate the Wasserstein Distance.
Relying solely on adversarial feedback is insufficient for the generator to capture the conditional
distribution of data. To improve the similarity between two distributions more efficiently, we
introduce two additional losses on the generator to discipline the learning. Firstly, the supervised
loss ℒ𝒮 is to directly describe the 𝐿1 distance between 𝑐 and 𝑐̂:
ℒ𝒮 = 𝔼𝑐̂∼𝑃𝑔,𝑐∼𝑃𝑟
[∥ 𝑐 − 𝑐̂ ∥] (14)
Secondly, from the perspective of stock movement, the long-term stock market shows an
increasing trend, hence the generator tends to construct sequences of higher return even with the
detection of short-term fluctuations, which leads to the abnormally high false positive ratio in the
classification performance. To address this problem, we design a weighted loss ℒ𝑊 which
represents the error rate with more focus on False Positive (FP) samples than False Negative (FN)
samples:
ℒ𝑊 =
1
𝑞
(λ1 ∑ 𝐼(𝑦𝑡 < 0, 𝑦𝑡
̂ > 0)
𝑞
𝑡=1
+ λ2 ∑ 𝐼(𝑦𝑡 > 0, 𝑦𝑡
̂ < 0)
𝑞
𝑖=1
) (15)
λ1 > λ2, λ1 + λ2 = 1, λ1 ∈ (0,1), λ2 ∈ (0,1)
where 𝐼(⋅) is the indicator function, the movement𝑦𝑡 = 𝑠𝑖𝑔𝑛(𝑐𝑡) is derived from the return,
similar with the predicted movement 𝑦
̂𝑡 = 𝑠𝑖𝑔𝑛(𝑐̂𝑡), 𝜆1 and 𝜆2 are hyper-parameters to balance
the error, the constraint 𝜆1 > 𝜆2 guarantees more penalty on FP ratio.
By integrating with the additional losses in adversarial feedback, the generator and critic
networks are trained simultaneously as follows:
ℒ𝒢 = −𝔼𝑐̂∼𝑃𝑔
[𝑓
𝑐(𝑐̂)] + 𝛾1ℒ𝒮 + 𝛾2ℒ𝑊 (16)
ℒ𝑓𝑐
= −𝔼𝑐∼𝑃𝑟
[𝑓
𝑐(𝑐)] + 𝔼𝑐̂∼𝑃𝑔
[𝑓
𝑐(𝑐̂)] (17)
where ℒ𝒢 and ℒ𝑓𝑐
are the objective functions for 𝒢 and 𝑓
𝑐 respectively, 𝛾1 ≥ 0and 𝛾2 ≥ 0 are
penalty parameters on supervised loss and weighted loss respectively.
To guarantee the Lipschitz constraint on the critic, it is necessary to force the weights of the critic
to lie in a compact space. In the training process, we clip the weights to a fixed interval [−𝛼, 𝛼]
after each gradient update.
4.4. Training Process
Figure 2 details the learning process for the generator𝒢 and critic 𝑓
𝑐 in the IndexGAN. Denote
𝒟𝑡𝑟𝑎𝑖𝑛 as the training data containing input features 𝑋 and real sequence 𝑐, and denote 𝒟𝑏𝑎𝑡𝑐ℎ as
a subset of 𝒟𝑡𝑟𝑎𝑖𝑛 for the training in each batch. We first initialize 𝒢 and 𝑓
𝑐. In each training
iteration, the prior noise 𝑧 sampled from Gaussian distribution and input features are fed into 𝒢 to
generate predicted sequence 𝑐̂, then critic estimates Wasserstein distance between real and fake
sequences. We compute the gradient from critic loss ℒ𝑓𝑐
and update parameters 𝛩𝑐 of 𝑓
𝑐 (line 8-9),
also clip all weights to accelerate them till optimality. We keep training critic for 𝑛𝑐𝑟𝑖𝑡𝑖𝑐 times,
then we focus on generator training. The reason for this iterative process is that gradient of
Wasserstein becomes reliable as the number of iterations for critic increases. Usually, 𝑛𝑐𝑟𝑖𝑡𝑖𝑐 is

20
set to 5. For generator training, predicted sequences are generated by 𝒢 and fed into 𝑓
𝑐. Then ℒ𝒢
is calculated by Eq. 14-16 and parameters Θ𝑔 of 𝒢 is updated (lines 15-16). The number of
epochs is set to 80. After the training in an adversarial manner, both 𝒢 and 𝑓
𝑐. with well-trained
parameters are converged and used for deployment to real-world data.
Figure 2 Training Process of IndexGAN
4.5. Deployment
Once the training is completed, the model is ready to be deployed. Recalling that generated
sequence from our proposed architecture covers 𝑞 steps ahead, the predicted price can be at either
location from 1 to 𝑞 in the generated sequence. Thus, we propose a rolling strategy for
deployment. To predict the target at time 𝑡′
, the model is repeated for 𝑞 rounds. For each round 𝑗,
where 𝑗 = 1, … , 𝑞, the historical features in the time interval [𝑡′
− 𝑤 + 1 − 𝑗, 𝑡′
− 𝑗]is learned to
generate sequence at time [𝑡′
+ 1 − 𝑗, 𝑡′
− 𝑗 + 𝑞] where 𝑐̂𝑡
𝑗
is the predicted return at 𝑗𝑡ℎ
position.
Following this method and rolling the window forward, we obtain a sequence of prediction
{𝑐̂𝑡
𝑗
}𝑗=1
𝑞
with dynamic change. We take the average of the sequence to obtain the final predicted
return 𝑐̂𝑡′ whose direction is the final predicted movement 𝑦
̂𝑡′ at time 𝑡′. The rolling strategy
considers the dynamic change of prediction since the trend of the stock market can be rapidly
influenced by the various factors discussed earlier. Meanwhile, the averaged predicted return
mitigates the noise caused by the uncertainty of the financial market. We expect the predicted
horizon 𝑞 to be 5 which can generate the best result considering every week has five trading days,
and our model evaluation verifies this conjecture.

21
5. MODEL EVALUATION
To evaluate the performance of the proposed model, extensive experiments are carried out with
real-world stock markets and news data. We will first introduce implementation details including
data description, baseline models, evaluation metrics, and parameter settings. Then we will show
the empirical results of compared models, forecasting horizon analysis, and ablation study to
demonstrate the effectiveness and generality of the proposed model, IndexGAN. The code is
publicly available1
.
5.1. Datasets
Stock datasets contain the historical daily stock prices in Dow Jones Industrial Average (DJIA)
and S&P 500 from Aug-08-2008 to Jul-01-2016, downloaded directly from Yahoo Finance. DJIA
is one of the oldest equity indices, tracking merely 30 of the most highly capitalized and
influential companies in major sectors of the U.S. market. S&P 500 is one of the most commonly
followed indices and is composed of the 500 largest publicly traded companies. Both indices are
used as reliable benchmarks of the economy. The chosen period covers a financial crisis and its
following recovery period, which provides stability to our method. Volatility indices of the same
period, VIX for S&P 500 and VXD for DJIA, are retrieved from Yahoo Finance. Both price and
volatility features 𝑠𝑡 are converted to return calculated by 𝑠𝑡/𝑠𝑡−1 − 1. The employed technical
indicators are calculated given the retrieved price data. The details are shown in Table 1.
News data is downloaded from Kaggle2
and is publicly available. It is crawled from Reddit
World News Channel, one of the largest social news aggregation websites. Content such as text
posts, images, links, and videos are uploaded by users to the site and voted up or down by other
users. The news and posts are related to popular events in various industry sectors in the world,
presenting an overall attitude with a high potential to influence the market. News data contains
the top 25 headlines ranked by users' votes for a single day. The headlines are processed by
removing stop words, punctuation, and digits from the raw text.
To tune the parameters, both stock and news data are split into training, validation, and testing
sets approximately as the ratio of 8:1:1 in chronological order to avoid data leakage, as shown in
Table 2. To capture the dependency of time series, a lag window with a size of 𝑤 time steps is
moved along both stock and news data to construct sets of sequences.
Table 2. Statistics of stocks and news datasets.
Dataset Time spans
Training 2008/09-2014/12
Validation 2014/12-2015/09
Testing 2015/09-2016/06
1
https://github.com/JingyiGu/IndexGAN
2
https://www.kaggle.com/aaron7sun/stocknews

22
5.2. Baselines
We compare the proposed IndexGAN with the following methods:
 MA [49] is one of the most popular indicators used by momentum traders. 5-day EMA is
chosen.
 ARIMA [50] is a widely used statistical model. In our implementation, the parameters
are determined automatically by the package 'pmdarima' in Python, regarding to the
lowest AIC.
 DA-RNN [19] integrates the attention mechanism with LSTM to extract input features
and relevant hidden states.
 StockNet [51] employs Variational Auto-Encoder (VAE) on tweets and historical prices
to encode stock movements as probabilistic vectors.
 A-LSTM [20] leverages adversarial training to add simulated perturbations on stock data
to improve the robustness of stock predictions.
 Trans [22] is the previous state-of-the-art model enhancing Transformer with trading gap
splitter and gaussian prior to predicting stock movement.
 DCGAN [52] is a powerful GAN that uses convolutional-transpose layers in the
generator and convolutional layers in the discriminator.
 GAN-FD [37] uses LSTM and convolutional layers to design a Vanilla GAN to predict
stock pricing one step ahead.
 S-GAN [45] employs FinBERT on sentiment scores and uses Vanilla GAN with
convolutional layers and GRU to predict pricing.
Note that MA and ARIMA use close prices only. DA-RNN, Adv-LSTM, Trans, DCGAN, and
GAN-FD use market-based prices only, following their architecture designs. StockNet and S-
GAN use both market-based features and news data.
Table 3. Parameters in IndexGAN.
Parameters DJI SPX
Batch size 32
Learning rate of generator 0.0001
Learning rate of critic 0.00005
GloVe embedding size 𝑚 50
Weight clipping threshold 𝛼 0.01
Trade-off parameter 𝜆1 0.8
Trade-off parameter 𝜆2 0.2
Trade-off parameter 𝛾1 10
Trade-off parameter 𝛾2 3
Number of Epochs 100 80
Latent word vector size 𝑔 6 3
Hidden size in Encoder 100 100
Hidden size in Decoder 200 50
5.3. Parameter Settings
The proposed model IndexGAN is implemented in PyTorch and hyper-parameters are tuned
based on the validation part. The length of historical sequence 𝑤 is set to 35 and future steps 𝑞
for prediction is set to 5. We use RMSprop as an optimizer. The details of the parameters in

23
IndexGAN are shown in Table 3. For other baselines, we follow default hyper-parameter settings
for implementation.
5.4. Evaluation Metrics
We evaluate the performance of all methods by three measures: Accuracy, F1 score, and
Matthews Correlation Coefficient (MCC). Accuracy is the proportion of correctly predicted
samples in all samples. F1 score incorporates both recall and precision. MCC takes all
components of the confusion matrix into consideration, while the true negative is ignored in F1
score. Therefore, MCC is in a more balanced manner than F1 score. A low value of MCC close to
0 indicates that classes in prediction are highly imbalanced.
5.5. Experiment Results
5.5.1. Comparison Results with Baselines
Table 4 presents the numerical performance of our IndexGAN and baselines on both data for
trend prediction, from which we make the following observations:
1. The proposed IndexGAN achieves the best results on both indices regarding accuracy and
MCC. On both data sets, IndexGAN achieves an accuracy of over 60% and MCC of
around 0.200, which justifies the predictive power of our model. The previous state-of-
the-art model Trans achieves the second-best performance but does not outperform
IndexGAN on MCC. Hence, IndexGAN is proven to surpass all baselines, including Trans,
on stock movement prediction tasks. In addition, the improvement in MCC by IndexGAN
is more significant than accuracy, which illustrates that prediction from IndexGAN is in a
more balanced manner.
2. Among the baselines, there exists an obvious gap between traditional time series models
(MA, ARIMA) modeling on close price only and neural networks (DA-RNN, StockNet
Adv-LSTM, Trans) modeling on all market-based features, demonstrating the impact of
additional features and the superior ability of neural networks in capturing complex
temporal dynamics for financial data. StockNet uses VAE, a generative model as well,
which maps features from high dimensional space to low dimension space and infers the
posterior distribution. Its performance is worse than IndexGAN. We postulate the reason
is that it cannot learn posterior distribution well due to the large noise of stock prices. It
reveals that GAN performs better in stock prediction tasks.
3. IndexGAN surpasses all GAN-based approaches. DCGAN is designed for image
generation and exploits convolution layers. Since patterns in time series are not similar as
in images, this architecture fails to capture temporal dependencies of stock features.
Hence, GANs for image generation is not suitable for time series data. Moreover, fake
and real sequences fed into the discriminator in GAN-FD are the same except for the last
element, which is challenging for the discriminator to differentiate. The low values of
MCC on both indices show that GAN-FD does not perform well on movement prediction.
Furthermore, S-GAN outperforms other GAN-based baselines due to its sentiment
analysis. However, the sentiment score from FinBERT in S-GAN falls into three
categories (positive, neutral, and negative). The complex pattern of public emotions is
hard to capture by such a simple sentiment score. The results of MCC illustrate that
Vanilla GAN implemented in S-GAN prevents the high diversity of prediction due to the
mode collapse problem. Therefore S-GAN falls behind IndexGAN. The proposed
structure of IndexGAN achieves two times of MCC increase, in general, because it
implements the concept of WGAN and multi-step prediction with horizon-wise
information.

24
Table 4. Performance comparison based on the average and standard deviation of 10 runs. The top two
baselines are traditional methods. The middle four models are recent state-of-the-art benchmarks. The
bottom three are based on GANs.
Model
DJIA S&P500
Acc (%) F1 MCC Acc (%) F1 MCC
EMA 51.83 0.483 0.058 53.93 0.511 0.087
ARIMA 53.15 0.479 0.091 50.53 0.475 0.085
DA-RNN 56.23 ±1.12 0.642 ±0.007 0.104 ±0.021 56.71 ±1.53 0.513 ±0.010 0.092 ±0.013
StockNet 57.90 ±0.84 0.632 ±0.008 0.111 ±0.031 55.37 ±1.35 0.585 ±0.009 0.091 ±0.019
A-LSTM 58.12 ±1.05 0.674 ±0.010 0.145 ±0.018 57.13 ±1.20 0.596 ±0.011 0.143 ±0.022
Trans 58.87 ±1.01 0.676 ±0.009 0.179 ±0.027 58.30 ±1.34 0.610 ±0.011 0.176 ±0.031
DCGAN 56.79 ±0.09 0.622 ±0.003 0.082 ±0.014 55.85 ±1.01 0.495 ±0.008 0.121 ±0.010
GAN-FD 57.02 ±1.09 0.607 ±0.011 0.094 ±0.023 56.93 ±1.15 0.591 ±0.012 0.105 ±0.024
S-GAN 57.93 ±1.08 0.632 ±0.010 0.099 ±0.020 57.42 ±1.07 0.601 ±0.012 0.112 ±0.018
IndexGAN 60.85 ±0.95 0.713 ±0.006 0.208 ±0.024 60.00 ±1.37 0.616 ±0.014 0.199 ±0.027
5.5.2. Forecasting Horizon Analysis
We investigate the effects of the prediction horizon 𝑞 on the proposed IndexGAN. Figure 3 shows
the change of accuracy and MCC on DJIA and SPX. As 𝑞 goes up from 2 to 5, the performance
of the model shows a rising trend, since the critic can differentiate real and fake sequences as
their length becomes longer. Both accuracy and MCC achieve the highest value when 𝑞 is 5.
However, as the length of the predicted sequence exceeds 5 days and continues increasing to 10
days, the quality of the model degrades. Thisreveals that it becomes harder to capture movement
patterns if the forecasting length is too long. Hence, there exists a trade-off between the
uncertainty of the stock market and the ability of GAN. Choosing the appropriate prediction
horizon is critical to achieving significant performance. The experiment conducted on various
lengths justifies our conjecture that 5 days forward is most effective because it is a normal and
popular short trading period.
Figure 3 Influence of different settings on predicted horizon
5.5.3. Ablation Study
We conduct experiments on ablation variants of IndexGAN to fully examine the predictive power
gained from each component:
 IndexGAN-Block: Blocks in news context learning are removed. The GloVe embedding
matrix for each headline is flattened and concatenated with price features as input to be
fed into seq2seq learning.

25
 IndexGAN-News: News headlines are removed from input features. The noise and
market-based features are concatenated together and fed into attentive seq2seq learning
directly. Thus, news context learning is removed as well.
 IndexGAN-Vol: The volatility index is removed from input features.
 IndexGAN -WDist: Wasserstein distance is removed and Vanilla GAN is implemented.
 IndexGAN-Attn: The attention layer in seq2seq learning is removed. The output from
GRU in the encoder is directly presented as input to the decoder.
 IndexGAN-Loss: ℒ𝒮 and ℒ𝑊 are removed and only adversarial loss is used.
Figure 4 exhibits the performance of variants. Firstly, IndexGAN-News, IndexGAN-Block, and
IndexGAN-FC are employed to test the importance of news context learning. With market-based
features only, the performance of IndexGAN-News decreases significantly. It validates that news
reflects the information in the stock market and can efficiently boost performance. On the other
hand, even with news, the performance of IndexGAN-Block is not ideal. It illustrates our notion
that directly passing the embedding matrix into the seq2seq model is likely to mask the impact of
price. When the embedding size is thousands of times greater than the price, the encoder and
decoder concentrate on embedding entirely. This phenomenon reveals that merely depending on
the news is challenging, and both price and news features are necessary for stock movement
prediction. Compared with a single fully connected layer, the blocks containing nonlinear layers
bring approximately a 3% accuracy increase on the S&P 500 and DJIA, and almost 0.05 of MCC
increase on the S&P 500. This phenomenon justifies that nonlinear layers in the designed blocks
can boost the predictive power efficiently. With the help of blocks that fully investigate the news
context and shrink the size of the embedding matrix gradually, IndexGAN achieves the highest
results.
Figure 4Ablation study results on Accuracy (top) and MCC (bottom) for comparing variants of IndexGAN.
For example, -Attn represents IndexGAN without attention in seq2seq learning.
Secondly, IndexGAN-WDist does not perform well, especially MCC is close to 0 which is even
worse than EMA, revealing low diversity of predicted movements and the mode collapse
problem of IndexGAN-WDist. Simply leveraging Vanilla GAN cannot capture dynamics for
movement prediction. It's necessary to employ the concept of Wasserstein distance to measure
the similarity of prediction and ground truth, which improves the quality of the model
significantly by accelerating the speed of convergence and avoiding mode collapse.
Thirdly, without the volatility index as a feature, the accuracy of Index-Vol on the S&P 500
decreases by 2.5% which is 1% more than the drop in accuracy on DJIA. The results indicate the
necessity of a volatility index, especially for VIX which depends on more options than VDX.
Moreover, by comparing the IndexGAN-Attn and IndexGAN-Loss with IndexGAN, the attention

26
mechanism and two additional losses on the generator enhance the predictive power by over a 2%
accuracy increase.
6. CONCLUSION
In this paper, we present a new formulation for predicting stock movement. This work is first to
propose the use of Wasserstein Generative Adversarial Networks for multi-step stock movement
prediction which includes particularly crafted designs for inherent characteristics of the stock
market. Specifically, in the generator, IndexGAN first exploits prior noise to improve the
robustness of the model, implements news context learning to explore embeddings from
emerging news, and employs volatility index as one of the impacting factors. Then an attentive
encoder-decoder architecture is designed for seq2seq learning to capture temporal dynamics and
interactions between market-based features and word representations. The critic is used for
approximating Wasserstein distance between the actual and predicted sequences. Moreover,
IndexGAN adds two additional losses for the generator to constrain similarity efficiently and
avoid biased movement during optimization. Also, to mitigate uncertainty, a rolling strategy for
deployment in employed. Experiments conducted on DJIA and S&P 500 show the significant
improvement of IndexGAN and the effectiveness of multi-step prediction over other state-of-art
methods. The ablation study verifies the significant advantage of each component in the proposed
IndexGAN as well. Given this, in future research, we will extend our work to other chaotic time
series where it is difficult to predict future performance from past data, especially those areas
driven by human factors (i.e., commerce market).
REFERENCES
[1] J. Bollen, H. Mao, and X. Zeng, “Twitter mood predicts the stock market,” Journal of computational
science, vol. 2, no. 1, pp. 1–8, 2011.
[2] R. Aguilar-Rivera, M. Valenzuela-Rend´on, and J. Rodr´ıguez-Ortiz, “Genetic algorithms and
darwinian approaches in financial applications: A survey,” Expert Systems with Applications, vol. 42,
no. 21, pp. 7684–7697, 2015.
[3] B. G. Malkiel, A random walk down Wall Street: including a life-cycle guide to personal investing.
WW Norton & Company, 1999.
[4] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y.
Bengio, “Generative adversarial nets,” Advances in neural information processing systems, vol. 27,
2014.
[5] X. Zhou, Z. Pan, G. Hu, S. Tang, and C. Zhao, “Stock market prediction on high-frequency data
using generative adversarial nets,” Mathematical Problems in Engineering, vol. 2018, 2018.
[6] M. Faraz and H. Khaloozadeh, “Multi-step-ahead stock market prediction based on least squares
generative adversarial network,” in 2020 28th Iranian Conference on Electrical Engineering (ICEE).
IEEE, 2020, pp. 1–6.
[7] A. Srivastava, L. Valkov, C. Russell, M. U. Gutmann, and C. Sutton, “Veegan: Reducing mode
collapse in gans using implicit variational learning,” in Proceedings of the 31st International
Conference on Neural Information Processing Systems, 2017, pp. 3310–3320.
[8] X. Ding, Y. Zhang, T. Liu, and J. Duan, “Deep learning for eventdriven stock prediction,” in Twenty-
fourth international joint conference on artificial intelligence, 2015.
[9] J. M. Poterba and L. H. Summers, “The persistence of volatility and stock market fluctuations,”
National Bureau of Economic Research, Tech. Rep., 1984.
[10] K. R. French, G. W. Schwert, and R. F. Stambaugh, “Expected stock returns and volatility,” Journal
of financial Economics, vol. 19, no. 1, pp. 3–29, 1987.
[11] G. W. Schwert, “Stock market volatility,” Financial analysts journal, vol. 46, no. 3, pp. 23–34, 1990.

27
[12] I. K. Nti, A. F. Adekoya, and B. A. Weyori, “A systematic review of fundamental and technical
analysis of stock market predictions,” Artificial Intelligence Review, pp. 1–51, 2019.
[13] C. A. Assis, A. C. Pereira, E. G. Carrano, R. Ramos, and W. Dias, “Restricted boltzmann machines
for the prediction of trends in financial time series,” in 2018 International Joint Conference on
Neural Networks (IJCNN). IEEE, 2018, pp. 1–8.
[14] L.-C. Cheng, Y.-H. Huang, and M.-E. Wu, “Applied attention-based lstm neural networks in stock
prediction,” in 2018 IEEE International Conference on Big Data (Big Data). IEEE, 2018, pp. 4716–
4718.
[15] O. B. Sezer and A. M. Ozbayoglu, “Algorithmic financial trading with deep convolutional neural
networks: Time series to image conversion approach,” Applied Soft Computing, vol. 70, pp. 525–538,
2018.
[16] R. J. Hyndman and G. Athanasopoulos, Forecasting: principles and practice. OTexts, 2018.
[17] T.Bollerslev, “Generalized autoregressive conditional heteroskedasticity,” Journal of econometrics,
vol. 31, no. 3, pp. 307–327, 1986.
[18] W. Yao, J. Gu, W. Du, F. P. Deek, and G. Wang, “Adpp: A novel anomalydetection and privacy-
preserving framework using blockchain and neuralnetworks in tokenomics,” International Journal of
Artificial Intelligence & Applications, vol. 13, no. 6, pp. 17–32, 2022
[19] Y. Qin, D. Song, H. Chen, W. Cheng, G. Jiang, and G. Cottrell, “A dualstage attention-based
recurrent neural network for time series prediction,” arXiv preprint arXiv:1704.02971, 2017.
[20] F. Feng, H. Chen, X. He, J. Ding, M. Sun, and T.-S. Chua, “Enhancing stock movement prediction
with adversarial training,” arXiv preprint arXiv:1810.09936, 2018.
[21] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural computation, vol. 9, no. 8, pp.
1735–1780, 1997.
[22] Q. Ding, S. Wu, H. Sun, J. Guo, and J. Guo, “Hierarchical multi-scale gaussian transformer for stock
movement prediction.” in IJCAI, 2020, pp. 4640–4646.
[23] X. Li, Y. Li, H. Yang, L. Yang, and X.-Y. Liu, “Dp-lstm: Differential privacy-inspired lstm for stock
prediction using financial news,” arXiv preprint arXiv:1912.10806, 2019.
[24] X. Zhang, H. Fuehres, and P. A. Gloor, “Predicting stock market indicators through twitter “i hope it
is not as bad as i fear”,” Procedia-Social and Behavioral Sciences, vol. 26, pp. 55–62, 2011.
[25] E. J. De Fortuny, T. De Smedt, D. Martens, and W. Daelemans, “Evaluating and understanding text-
based stock price prediction models,” Information Processing & Management, vol. 50, no. 2, pp.
426–441, 2014.
[26] B. Zhao, Y. He, C. Yuan, and Y. Huang, “Stock market prediction exploiting microblog sentiment
analysis,” in 2016 International Joint Conference on Neural Networks (IJCNN). IEEE, 2016, pp.
4482–4488.
[27] J. Tang and X. Chen, “Stock market prediction based on historic prices and news titles,” in
Proceedings of the 2018 International Conference on Machine Learning Technologies, 2018, pp. 29–
34.
[28] J. B. Nascimento and M. Cristo, “The impact of structured event embeddings on scalable stock
forecasting models,” in Proceedings of the 21st Brazilian Symposium on Multimedia and the Web,
2015, pp. 121–124.
[29] Y. Liu, Q. Zeng, H. Yang, and A. Carrio, “Stock price movement prediction from financial news with
deep learning and knowledge graph embedding,” in Pacific rim Knowledge acquisition workshop.
Springer, 2018, pp. 102–113.
[30] C.-Y. Lee and V.-W. Soo, “Predict stock price with financial news based on recurrent convolutional
neural networks,” in 2017 conference on technologies and applications of artificial intelligence
(TAAI). IEEE, 2017, pp. 160–165.
[31] D. L. Minh, A. Sadeghi-Niaraki, H. D. Huy, K. Min, and H. Moon, “Deep learning approach for
short-term stock trends prediction based on twostream gated recurrent unit network,” Ieee Access, vol.
6, pp. 55392– 55404, 2018.
[32] W. Li, R. Bao, K. Harimoto, D. Chen, J. Xu, and Q. Su, “Modeling the stock relation with graph
network for overnight stock movement prediction,” in Proceedings of the Twenty-Ninth International
Conference on International Joint Conferences on Artificial Intelligence, 2021, pp. 4541–4547.
[33] M. Saito, E. Matsumoto, and S. Saito, “Temporal generative adversarial nets with singular value
clipping,” in Proceedings of the IEEE international conference on computer vision, 2017, pp. 2830–
2839.

28
[34] C. Esteban, S. L. Hyland, and G. R¨atsch, “Real-valued (medical) time series generation with
recurrent conditional gans,” arXiv preprint arXiv:1706.02633, 2017.
[35] Y. Zhang, Z. Gan, and L. Carin, “Generating text via adversarial training,” in NIPS workshop on
Adversarial Training, vol. 21. academia. edu, 2016, pp. 21–32.
[36] A. Gupta and J. Zou, “Feedback gan for dna optimizes protein functions,” Nature Machine
Intelligence, vol. 1, no. 2, pp. 105–111, 2019.
[37] S. Takahashi, Y. Chen, and K. Tanaka-Ishii, “Modeling financial timeseries with generative
adversarial networks,” Physica A: Statistical Mechanics and its Applications, vol. 527, p. 121261,
2019.
[38] H.-W. Dong, W.-Y. Hsiao, L.-C. Yang, and Y.-H. Yang, “Musegan: Multi-track sequential
generative adversarial networks for symbolic music generation and accompaniment,” in Thirty-
Second AAAI Conference onArtificial Intelligence, 2018.
[39] L. Yu, W. Zhang, J. Wang, and Y. Yu, “Seqgan: Sequence generative adversarial nets with policy
gradient,” in Proceedings of the AAAI conference on artificial intelligence, vol. 31, no. 1, 2017.
[40] J. Yoon, D. Jarrett, and M. Van der Schaar, “Time-series generative adversarial networks,” 2019.
[41] M. Wiese, R. Knobloch, R. Korn, and P. Kretschmer, “Quant gans: Deep generation of financial time
series,” Quantitative Finance, vol. 20, no. 9, pp. 1419–1440, 2020.
[42] Y. Xu, Z. Zhang, L. You, J. Liu, Z. Fan, and X. Zhou, “scigans: single-cell rna-seq imputation using
generative adversarial networks,” Nucleic acids research, vol. 48, no. 15, pp. e85–e85, 2020.
[43] L. Wu, Y. Xia, F. Tian, L. Zhao, T. Qin, J. Lai, and T.-Y. Liu, “Adversarial neural machine
translation,” in Asian Conference on Machine Learning. PMLR, 2018, pp. 534–549.
[44] K. Zhang, G. Zhong, J. Dong, S. Wang, and Y. Wang, “Stock market prediction based on generative
adversarial network,” Procedia computer science, vol. 147, pp. 400–406, 2019.
[45] P. Sonkiya, V. Bajpai, and A. Bansal, “Stock price prediction using bert and gan,” arXiv preprint
arXiv:2107.09055, 2021.
[46] M. Mirza and S. Osindero, “Conditional generative adversarial nets,” arXiv preprint arXiv:1411.1784,
2014.
[47] M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein generative adversarial networks,” in
International conference on machine learning. PMLR, 2017, pp. 214–223.
[48] Y. Rubner, C. Tomasi, and L. J. Guibas, “The earth mover’s distance as a metric for image retrieval,”
International journal of computer vision, vol. 40, no. 2, pp. 99–121, 2000.
[49] F. Klinker, “Exponential moving average versus moving exponential average,” Mathematische
Semesterberichte, vol. 58, no. 1, pp. 97–107, 2011.
[50] S. Makridakis and M. Hibon, “Arma models and the box–jenkins methodology,” Journal of
forecasting, vol. 16, no. 3, pp. 147–163, 1997.
[51] Y. Xu and S. B. Cohen, “Stock movement prediction from tweets and historical prices,” in
Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1:
Long Papers), 2018, pp. 1970–1979.
[52] A. Radford, L. Metz, and S. Chintala, “Unsupervised representation learning with deep convolutional
generative adversarial networks,” arXiv preprint arXiv:1511.06434, 2015.

STOCK BROAD-INDEX TREND PATTERNS LEARNING VIA DOMAIN KNOWLEDGE INFORMED GENERATIVE NETWORK

More Related Content

Similar to STOCK BROAD-INDEX TREND PATTERNS LEARNING VIA DOMAIN KNOWLEDGE INFORMED GENERATIVE NETWORK (20)

More from gerogepatton (20)

Recently uploaded (20)

STOCK BROAD-INDEX TREND PATTERNS LEARNING VIA DOMAIN KNOWLEDGE INFORMED GENERATIVE NETWORK